Re: [PATCH 3/4] proc: Point /proc/net at /proc/thread-self/net instead of /proc/self/net

[Al Viro]

On Thu, Sep 29, 2022 at 10:21:04PM +0100, Al Viro wrote:
> On Thu, Sep 29, 2022 at 02:13:57PM -0700, Linus Torvalds wrote:
> > On Thu, Sep 29, 2022 at 12:34 PM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > Apparmor takes mount+dentry and turns that into pathname.  Then acts
> > > upon the resulting string.  *AFTER* the original had been resolved.
> > 
> > Ok. So it would have to act like a bind mount.
> > 
> > Which is probably not too bad.
> > 
> > In fact, maybe it would be ok for this to act like a hardlink and just
> > fill in the inode - not safe for a filesystem in general due to the
> > whole rename loop issue, but for /proc it might be fine?
> 
> _Which_ hardlink?
> 
> Linus, where in dentry tree would you want it to be seen?  Because
> apparmor profile wants /proc/net/dev to land at /proc/<pid>/net/dev
> and will fail with anything else.
> 
> Do you really want multiple dentries with the same name in the same
> parent, refering to different directory inodes with different contents?
> 
> And that's different inodes with different contents - David's complaint
> is precisely about seeing the same thing for all threads and apparmor
> issue is with *NOT* seeing each of those things at the same location.

Put it another way:

David:
	when I'm opening /proc/net/whatever, I want its contents to match
	this thread's netns, not that of some other thread.
dhclient+apparmor:
	whatever you get from /proc/net/dev, it would better be at
	/proc/<pid>/net/dev, no matter which thread you happen to be.

It's not that we want to see the same thing in several places; it's that
we want to see *different* things in the same place.  Opposite to what
hardlinks or bindings would be about.